This is v4.1 of the SFT training, focusing on data curation via vector search. Based on v3's lesson (score dropped from 0.751 to 0.726 with more data), we curated the training data.

Changes from v3

Parameter	v3	v4.1	Rationale
Dataset	Merged (8,541)	Curated k=50 (1,695)	Quality over quantity
MAX_SEQ_LEN	1024	1024	Same as v2/v3
Epochs	1	1	Same as v2/v3
Learning Rate	5e-6	5e-06	Same as v2/v3

Curated Dataset Details

Method: sentence-transformers + k-NN search for test-similar data
k: 50 neighbors per test sample
Total: 1,695 samples (71% of v4)

Source distribution: 2-2_5k_mix(46%), 2-1_3k_mix(35%), 2-3_hard_4k(19%)

Training Objective

This adapter is trained to improve structured output accuracy (JSON / YAML / XML / TOML / CSV) for the StructEval-T benchmark.

Loss is applied only to the final assistant output, while intermediate reasoning (Chain-of-Thought) is masked.

Training Configuration

Base model: Qwen/Qwen3-4B-Instruct-2507
Method: QLoRA (4-bit, Unsloth)
Max sequence length: 1024
Epochs: 1
Learning rate: 5e-06
Batch size: 2 (effective: 16)
Gradient accumulation: 8
LoRA: r=64, alpha=128
CoT masking: enabled (loss on final output only)

Usage

python

from transformers import AutoModelForCausalLM, AutoTokenizer
from peft import PeftModel
import torch

base = "Qwen/Qwen3-4B-Instruct-2507"
adapter = "your_id/qwen3-4b-structured-output-lora-v4.1"

tokenizer = AutoTokenizer.from_pretrained(base)
model = AutoModelForCausalLM.from_pretrained(
    base,
    torch_dtype=torch.float16,
    device_map="auto",
)
model = PeftModel.from_pretrained(model, adapter)

Sources & Terms (IMPORTANT)

Training data: merged dataset (1-1_512_v2 + 1-2_512_v4)

Dataset License: MIT License. This dataset is used and distributed under the terms of the MIT License. Compliance: Users must comply with the MIT license (including copyright notice) and the base model's original terms of use.

kmd2525/qwen3-4b-structured-output-lora-v4.1

qwen3-4b-structured-output-lora-v4.1

Version: v4.1 — Data Curation (k=50)